Dask client should be optional

Currently, a dask client / cluster connection is needed for fitting.

However, it should be possible to not use this but instead do the fitting in a the same thread (blocking)

ThreadPoolExecutor is an option but was found to be very slow for PyTorch fitting